Analysis of Aerial Images Using Deep Learning to Identify Critical Areas in Natural Disasters
Nidhya Shivakumar
Abstract
Climate change has led to a rise in the frequency and intensity of natural disasters, which, in turn, affect a wide swath of human-populated areas. The use of Unmanned Aerial Vehicles (UAVs) is on the rise in mapping out the disaster areas in their immediate aftermath. However, processing the vast amount of data obtained can take several hours to days, costing crucial time that could be used in saving lives and infrastructure. In this study, methods were developed to automate and accelerate the identification of areas that are in critical need of assistance. A Faster R-CNN object detection model was built to classify buildings into damaged and undamaged with 90% precision, and to further sub-classify them by the type of damage (undamaged, flood, rubble). The number of high quality labeled images required for training models was increased by 163% by developing an auto-label generation technique using weak supervision. Ensemble modeling further improved the recall of model predictions by 16%, and with higher prediction accuracy, when analyzing disasters not included in the training set. The utility of the model was demonstrated by using it to produce an annotated video of the 2021 tornado damage in Kentucky. A Mask R-CNN segmentation model had the best performance overall and identified undamaged roads with 100% precision and building damage with precision greater than 84% and with recall greater than 74% for all object types. This study demonstrates the power of deep learning in the processing of images from disaster-stricken areas to aid search and rescue efforts and significantly reduce disaster response times.